1 / 54

A Primal-Dual Approach for Online Problems

A Primal-Dual Approach for Online Problems. Nikhil Bansal. Online Algorithms. Input revealed in parts. Algorithm has no knowledge of future. Scheduling, Load Balancing, Routing, Caching, Finance, Machine Learning … Competitive ratio = Expected Competitive ratio =.

adanna
Download Presentation

A Primal-Dual Approach for Online Problems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Primal-Dual Approach for Online Problems Nikhil Bansal

  2. Online Algorithms Input revealed in parts. Algorithm has no knowledge of future. Scheduling, Load Balancing, Routing, Caching, Finance, Machine Learning … Competitive ratio = Expected Competitive ratio = Alternate view: game between algorithm and adversary

  3. Some classic problems

  4. The Ski Rental Problem • Buying costs $B. • Renting costs $1 per day. Problem: • Number of ski days is not known in advance. Goal: Minimize the total cost. Deterministic: 2 Randomized: e/(e-1) ¼ 1.58

  5. Online Virtual Circuit Routing Network graph G=(V, E) capacity function u: E Z+ Requests: ri = (si, ti) • Problem: Connect si to ti by a path, or reject the request. • Reserve one unit of bandwidth along the path. • No re-routing is allowed. • Load: ratio between reserved edge bandwidth and edge capacity. • Goal: Maximize the total throughput.

  6. s3 t2 s1 t1 t3 s2 1/5 2/5 3/5 Virtual Circuit Routing - Example Edge capacities: 5 Maximum Load: 0

  7. Virtual Circuit Routing Key decision: • Whether to choose request or not? • How to route request? O(log n)-congestion, O(1)-throughput [Awerbuch Azar Plotkin 90’s] Various other versions and tradeoffs. Main idea:Exponential penalty approach length (edge) = exp (congestion) Decisions based on length of shortest (si,ti) path Clever potential function analysis

  8. The Paging/Caching Problem Set of pages {1,2,…,n} , cache of size k<n. Request sequence of pages1, 6, 4, 1, 4, 7, 6, 1, 3, … a) If requested page already in cache, no penalty. b) Else, cache miss. Need to fetch page in cache (possibly) evicting some other page. Goal: Minimize the number of cache misses. Key Decision: Upon a request, which page to evacuate?

  9. Previous Results: Paging Paging (Deterministic) [Sleator Tarjan 85]: • Any det. algorithm ¸k-competitive. • LRU is k-competitive (also other algorithms) Paging (Randomized): • Rand. Marking O(log k)[Fiat, Karp, Luby, McGeoch, Sleator, Young 91]. • Lower bound Hk[Fiat et al. 91], tight results known.

  10. Do these problems have anything in common?

  11. An Abstract Online Problem min 3 x1 + 5 x2 + x3 + 4 x4 + … 2 x1 + x3 + x6 + … ¸ 3 x3 + x14 + x19 + … ¸ 8 x2 + 7 x4 + x12 + … ¸ 2 Goal: Find feasible solution x* with min cost. Requirements: 1) Upon arrival constraint must be satisfied 2) Cannot decrease a variable. Covering LP (non-negative entries)

  12. Example min x1 + x2 + … + xn x1 + x2 + x3 + … + xn¸ 1 x2 + x3 + … + xn¸ 1 x3 + … + xn¸ 1 … xn¸ 1 Online ¸ln n(1+1/2+ 1/3+ … + 1/n) Opt = 1( xn=1 suffices) Set all xi to 1/n Increase x2 ,x3,…,xn to 1/n-1 … Increase xn to 1

  13. The Dual Problem max 3 y1 + 5 y2 + y3 + 4 y4 + … 2 y1+ y2 + y3 + … · 3 y1 + y2 + 2 y3 + … · 8 y1 +7 y2 + y3 + … · 2 Goal: Find y* with max cost. Requirements: 1) Variables arrive sequentially 2) At step t, can only modify y(t) Packing LP (non-negative entries) All previous problems can be expressed as Covering/Packing LP

  14. Ski Rental – Integer Program Subject to: For each day i: (either buy or rent)

  15. Routing – Linear Program = Amount of bandwidth allocated for ri on path p s.t: For each ri: For each edge e: - Available paths to serve request ri

  16. Paging – Linear Program (i,2) (i,1) Time line Pg i’ Pg i Pg i Pg i’ Pg i’ Pg i t If interval not present, then cache miss. At any time t, can have at mostk such intervals.  at leastn-k intervals must be absent n: number of distinct pages x(i,j): How much interval (i,j) evacuated thus far Cost =ij x(i,j) i: i  pt x(i,r(i,t)) ¸ n-k 8 t 0 · x(i,j) · 1

  17. What can we say about the abstract problem ?

  18. General Covering/Packing Results For a {0,1} covering/packing matrix: [Buchbinder Naor 05] • Competitive ratio O(log D) • Can get e/e-1 for ski rental and other problems. (D – max number of non-zero entries in a constraint). Remarks: • Fractional solutions • Number of constraints/variables can be exponential. • There can be a tradeoff between the competitive ratio and the factor by which constraints are violated. Fractional solution ! randomized algorithm (online rounding)

  19. General Covering/Packing Results For a general covering/packing matrix [BN05] : Covering: • Competitive ratio O(log n) (n – number of variables). Packing: • Competitive ratio O(log n + log [a(max)/a(min)]) a(max), a(min) – max/min non-zero entry Remarks: • Results are tight. • Can add “box” constraints to covering LP (e.g. x · 1)

  20. Consequences The online covering LP problem (and its dual packing counterpart) is a powerful framework Ski-Rental, Adword auctions, Dynamic TCP acknowledgement, Online Routing, Load Balancing, Congestion Minimization, Caching, Online Matching, Online Graph Covering, Parking Permit Problem, … Routing: O( log n) congestion, 1 competitive on throughput Can incorporate fairness Awerbuch, Azar, Plotkin result obtained by derandomizing the scheme online by applying pessimistic estimators.

  21. web Consequences (Weighted Paging) • Each page i has a different fetching cost w(i). Main memory, disk, internet … Goal: Minimize the total cost of cache misses. O(log k) competitive algorithm [B., Buchbinder, Naor 07] Previously, o(k) known only for the case of 2 weights [Irani 02] O(log2 k) for Generalized Paging (arbitrary weights and sizes) [B., Buchbinder, Naor 08] Previously, o(k) known only for special cases. [Irani 97]

  22. Rest of the Talk • Overview of LP Duality, offline P-D technique • Derive Online Primal Dual (very natural) • Further Extensions

  23. Duality Min 3 x1 + 4 x2 x1 + x2 >= 3 x1 + 2 x2 >= 5 Want to convince someone that there is a solution of value 12. Easy, just demonstrate a solution, x2 = 3

  24. Duality Min 3 x1 + 4 x2 x1 + x2 >= 3 x1 + 2 x2 >= 5 Want to convince someone that there is no solution of value 10. How? 2 * first eqn + second eqn 3 x1 + 4 x2 >= 11 LP Duality Theorem: This seemingly ad hoc trick always works!

  25. LP Duality Min cj xj j aij xj¸ bi So, for any y ¸ 0 satisfying i aij yi· cj for all i j xj cj¸i yi bi Equality when Complementary Slackness i.e. yi > 0 (only if corresponding primal constraint is tight) xi > 0 (only if corresponding dual constraint is tight) Linear combination i yij aij xj¸i yi bi j xj ( i aij yi ) ¸i yi bi (y ¸ 0) Dual LP Dual cost

  26. Offline Primal-Dual Approach min cx max b y Ax ¸ b At y · c x ¸ 0 y ¸ 0 Generic Primal Dual Algorithm: 0) Start with x=0, y=0 (primal infeasible, dual feasible) 1) Increase dual and primal together, s.t. if dual cost increases by 1, primal increases by · c 2) If both dual and primal feasible ) c approximate solution

  27. Key Idea for Online Primal Dual Primal: Min i ci xi Dual Step t, newconstraint:New variable yt a1x1 + a2x2 + … + ajxj¸ bt + bt yt in dual objective How much:  xi ? yt! yt + 1 (additive update) primal cost = = Dual Cost dx/dy proportional to x so, x varies as exp(y)

  28. How to initialize A problem: dx/dy is proportional to x, but x=0 initially. So, x will remain equal to 0 ? Answer: Initialize to 1/n. When: Complementary slackness tells us that x > 0 only if dual constraint corresponding to x is tight. Set x=1/n when its dual constraint becomes tight.

  29. The Algorithm Min j cj xj j aij xj¸ bi On arrival of i-th constraint, Initializeyi=0 (dual var. for constraint) If current constraint unsatisfied, gradually increase yi If xj =0, set xj = 1/n when i aij yi = cj else update xj as 1/n ¢ exp( (i aij yi / cj) - 1 ) Max i bi yi i aij yi· cj 1) Primal Cost · Dual Cost 2) Dual solution violated by at most O(log n) factor.

  30. Example: Caching xp: fraction of p missing from cache 1 1/k 0 Corresponding Dual constraint Dual violated by O(log k) Dual is tight Page fully in cache (“marked”) Page fully evacuated Page is “unmarked”

  31. Part 2: Rounding Primal dual technique gives fractional solution. Problem specific rounding/interpretation: 1) Easy for ski rental (value of x, is prob. of buying by then) 2) Routing: Can derandomize online using pessimistic estimator or other techniques 3) Caching (tricky): Gives probability distribution on pages, Actually want probability distribution on cache states.

  32. Beyond Packing/Covering LPs

  33. Extended Framework Limitations of current framework 1. Only covering or packing LP 2. Variables can only increase. Cannot impose: a ¸ b or a ¸ b1 – b2 Problem with monotonicity: Predicting with Experts: Do as well as best expert in hindsight n experts: Each day, predict rain or shine. Online · Best expert (1+ ) + O(log n)/ (low regret) In any LP, xi,t = Prob. of expert i at time t.

  34. New LP for weighted paging Variable yp,t: How much page p missing from cache at time t. pt = page requested at time t. Min p,t wp zp,t + t1¢ ypt,t p 2 S yp,t¸ |S|-k 8 S,t zp,t¸ yp,t – yp,t-18 p,t yp,t¸ 0 8 p,t The insights from previous approach can be used. Notably, multiplicative updates Solve finely competitive paging. [B., Buchbinder, Naor 10]

  35. K-Server Problem

  36. 3 1 2 Move Nearest Algorithm The k-server Problem • k servers lie in an n-point metric space. • Requests arrive at metric points. • To serve request: Need to move some server there. Goal: Minimize total distance traveled. Objective:Competitive ratio.

  37. . . . n 1 The Paging/Caching Problem K-server on the uniform metric. Server on location p = page p in cache

  38. Previous Results: Paging Paging (Deterministic) [Sleator Tarjan 85]: • Any deterministic algorithm >= k-competitive. • LRU is k-competitive (also other algorithms) Paging (Randomized): • Rand. Marking O(log k)[Fiat, Karp, Luby, McGeoch, Sleator, Young 91]. • Lower bound Hk[Fiat et al. 91], tight results known.

  39. K-server conjecture [Manasse-McGeoch-Sleator ’88]: There exists k competitive algorithm on any metric space. Initially no f(k) guarantee. Fiat-Rababi-Ravid’90: exp(k log k) … Koutsoupias-Papadimitriou’95:2k-1 Chrobak-Larmore’91: k for trees.

  40. Randomized k-server Conjecture There is an O(log k) competitive algorithm for any metric. Uniform Metric: log k Polylog for very special cases (uniform-like) Line: n2/3[Csaba-Lodha’06] exp(O(log n)1/2) [Bansal-Buchbinder-Naor’10] Depth 2-tree: No o(k) guarantee

  41. Result Thm [B.,Buchbinder,Madry,Naor 11]: There is an O(log2 k log3 n) competitive* algorithm for k-server on any metric with n points. Key Idea: Multiplicative Updates * Hiding some log log n terms

  42. Our Approach Hierarchically Separated Trees (HSTs)[Bartal 96]. Any Metric Allocation Problem (uniform metrics): [Cote-Meyerson-Poplawski’08] (decides how to distribute servers among children) O(log n) Allocation instances K-server on HST

  43. Outline • Introduction • Allocation Problem • Fractional Caching Algorithm • The final solution

  44. Allocation Problem Uniform Metric At each time t, request arrives at some location i Request = (ht(0),…,ht(k)) [monotone: h(0) ¸ h(1) … ¸ h(k)] Upon seeing request, can reallocate servers Hit cost = ht(ki) [ki : number of servers at i] Total cost = Hit cost + Move cost Eg: Paging = cost vectors (1,0,0,…,0) *Total servers k(t) can also change (let’s ignore this)

  45. Allocation to k-server Thm [Cote-Poplawski-Meyerson]: An online algorithm for allocation s.t. for any  > 0, i) Hit Cost (Alg) ·(1+) OPT ii) Move Cost (Alg) ·(e) OPT gives ¼O(d (1/d)) competitive k-server alg. on depth d HSTs d = log (aspect ratio) So,  = poly(1/) polylog(k,n) suffices *HSTs need some well-separatedness *Later, we do tricks to remove dependence on aspect ratio We do not know how to obtain such an algorithm.

  46. Fractional Allocation Problem xi,j: prob. of having j servers at location i (at time t) j xi,j = 1 (prob. distribution) ij j xi,j· k (global server bound) Cost: Hit cost with h(0),…,h(k) = j xi,j h(j) Moving  mass from (i,j) to (i,j’) costs  |j’-j| Surprisingly, fractional allocation does not give good randomized alg. for allocation problem.

  47. A gap example Allocation Problem on 2 points Left Right Requests alternate on locations. Left: (1,1,…,1,0) Right: (1,0,…,0,0) Any integral solution must pay (T) in T steps. Claim: Fractional Algorithm pays only T/(2k) . XL,0 = 1/k xL,k = 1-1/k XR,1 = 1 No move cost. Hit cost of 1/k on left requests.

  48. Fractional Algorithm Suffices Thm (Analog of Cote et al): Suffices to have fractional allocation algorithm with (1+,()) guarantee. Gives a fractional k-server algorithm on HST Thm (Rounding): Fractional k-server alg. on HSTs -> Randomized Alg. with O(1) loss. Thm (Frac. Allocation): Design a fractional allocation algorithm with (e) = O(log (k/)).

  49. Outline • Introduction • Allocation Problem • Fractional Caching Algorithm • The final Solution

  50. 0 1 0 1 0 1 0 1 P1,0 P1,1 Pn,0 Pn,1 p2,0 P2,1 … Pg n Pg 1 Pg 2 Fractional Paging Algorithm State: For each location i, we have pi,0 + pi,1=1 and i pi,1 =k. Say request at 1 arrives. Algorithm: Need to bring p1,0=1-p1,1 mass into p1,1. Rule: For each page i decrease pi,1/pi,0 +  ( = 1/k) Intuition: If pi,1 close to 1, be more conservative in evicting. Multiplicative Update: d(p) / (p)

More Related